Unsupervised Learning of Signal Strength Models for Device-Free Localization

Abstract

RSS-based device-free localization (DFL) systems make use of the received signal strength (RSS) changes in a network of static wireless nodes to locate and track people. Current DFL systems require calibration, which depending on the method and required accuracy, can be very expensive in terms of time and effort, making DFL system deployment and maintenance challenging. This paper implements unsupervised learning of signal strength models (UnLeSS), a Baum-Welch based method to learn the parameters of a hidden Markov model (HMM) for each link, including the RSS distribution during the no-crossing state and the crossing state. The system uses the HMM to estimate the probability of each link being in the crossed state. As a demonstration of its effectiveness, the per-link probability is used in a radio tomographic imaging algorithm to track the location of a person. Experiments are conducted in two different homes to determine the performance of UnLeSS. We demonstrate that our system is capable of estimating the crossing/no-crossing distribution with Kullback-Leibler divergence maximum of 1.43. UnLeSS is capable of tracking a person with high accuracy (0.66 m) without a calibration period.